1. technology and computing
2. programming languages
3. javascript

Stickler: Defending Against Malicious CDNs
in an Unmodified Browser
Amit Levy
Henry Corrigan-Gibbs
Dan Boneh
levya@cs.stanford.edu
Stanford University
Stanford, CA
henrycg@stanford.edu
Stanford University
Stanford, CA
dabo@cs.stanford.edu
Stanford University
Stanford, CA
Abstract—Website publishers can derive enormous performance benefits and cost savings by directing traffic to their
sites through content distribution networks (CDNs). However,
publishers who use CDNs today must trust their CDN not to
modify the site’s JavaScript, CSS, images or other media en
route to end users. A CDN that violates this trust could inject ads
into websites, downsample media to save bandwidth or, worse,
inject malicious JavaScript code to steal user secrets it could
not otherwise access. We present Stickler, a system for website
publishers that guarantees the end-to-end authenticity of content
served to end users while simultaneously allowing publishers
to reap the benefits of CDNs. Crucially, Stickler achieves these
guarantees without requiring modifications to the browser.
I.
I NTRODUCTION
Transport-layer encryption (e.g., as used in HTTPS) is
the industry standard for protecting the confidentiality and
authenticity of data in transit on the Web. If the end user
connects directly to a website publisher’s servers then HTTPS
is sufficient to authenticate the web content served over the
connection. However, when there is an intermediary, such as
a content distribution network (CDN), between the end user
and the publisher, the HTTPS connection terminates at the
CDN’s servers. In this circumstance, HTTPS ensures that the
connection to the CDN is authenticated, but it says nothing
about whether the CDN is serving the publisher’s intended
content. Essentially, the publisher must completely trust the
CDN to faithfully serve the site’s assets to its users.
But how much should website publishers trust the CDNs
that host their sites? Today, publishers have no choice but to
assume that CDNs are not modifying their sites’ JavaScript,
images, and other assets en route to end users. However, this
assumption is not always reasonable. For example, CDNs can
inject ads in order to increase revenue, may be compelled (e.g.,
by powerful governments) to modify JavaScript assets to leak
passwords, or might downsample image files to reduce their
own bandwidth costs. Moreover, while publishers might be
able to catch misbehaving CDNs by sampling CDN-served
content, publishers may be unable to catch CDNs that modify
content in targeted ways (e.g. only for certain IP addresses, in
certain countries or after certain times).
In the last few years, a small number of CDN providers
have emerged to provide content delivery services to an increasing number of websites—CloudFlare alone claims to host
content for over two million websites [1]. These “consumergrade” CDNs are much cheaper and have a much looser
relationship with their customers than do traditional CDNs.
Typically, website publishers sign up for an account with
low-cost CDNs via a click-through web interface and often
pay nothing for the service. Even though these CDNs have
a tenuous business relationship with the sites they host, the
sites’ publishers are implicitly delegating a huge amount of
trust to their CDNs: in at least one case, a no-cost CDN
was able to generate CA-signed HTTPS certificates for its
publishers’ domains without the publishers’ intervention [2].
Today, website publishers must weigh the uncertain risks of
using such a service against the undeniable benefits: better
availability, cheaper bandwidth, and “flash crowd” protection
[3]–[8].
In this paper, we describe Stickler: a system that guarantees
end-to-end integrity of website content in the face of malicious
CDNs, without modifying existing browsers or CDNs. Stickler
separates authenticity of the connection from authenticity of
the content by signing content directly with a private key that
the site owner never has to share with the CDN. In fact, if
the site consists entirely of static content, even the publisher’s
own web servers need not hold a copy of the publisher’s private
signing key. End users can verify all site content independently
of how they receive the pre-signed content, whether directly
from the web site’s servers, through a minimally trusted CDN
provider, or even via a peer-to-peer CDN like CoralCDN [3].
Unlike the proposed W3C Subresource Integrity mechanism (SRI) for protecting the integrity of CDN-served content [9], Stickler does not require modifying the browser.
Although browser support for integrity protection would be
ideal from a performance perspective, implementing and deploying these mechanisms in commodity browsers may take
years and may never reach all browser vendors and platforms.
For example, Server Name Indication (SNI) was introduced
in 2003 [10] and is widely used on servers to host multiple HTTPS sites on the same IP address, but did not gain
widespread adoption in clients until 2011 [11] and is still not
universally supported [12]. Stickler is an effective short-term
solution for browsers that will eventually support SRI and is
a useful long-term solution for browsers and platforms (e.g.,
old smartphones) that may never support SRI. We discuss SRI,
the Content Security Policy mechanism [13], and other related
work in more detail in Section VII.
This paper makes three main contributions:
1) Architecture. We introduce a CDN-compatible architecture for providing end-to-end web integrity with minimal
changes to existing infrastructure.
2) Implementation. We present Stickler, an implementation
of this architecture that works with existing servers,
CDNs, and web browsers.
3) Performance Evaluation. We evaluate the performance
of Stickler to demonstrate that it is practical and feasible
to deploy on today’s web.
The remainder of the paper is organized as follows:
Section II presents background on CDNs and their benefits.
Section III formalizes the threat model. Section IV presents an
architecture that preserves end-to-end integrity of content from
a malicious CDN. Section V describes our implementation of
Stickler. Section VI evaluates the performance overhead of
Stickler in the end user’s web browser. Section VII discusses
related work and Section VIII concludes.
II.
BACKGROUND
In this section, we introduce the terms used throughout this
paper and explain the technical environment in which a site
publisher would deploy Stickler.
A. Participants
The four parties relevant to a Stickler deployment are: the
site publisher, the content distribution network, the certificate
authority, and the end user.
Publisher. The site publisher is the entity that maintains
the website in question. For example, The New York Times
Company is the publisher of nytimes.com. Publishers use
Stickler to ensure that visitors to their site receive authentic
publisher-produced content.
Content Distribution Network (CDN). A publisher may
contract a content distribution network to serve the site’s static
web assets (images, videos, etc.) or cacheable dynamic content.
Publishers typically give their CDN complete control of a
specific domain on which the CDN serves the publisher’s
content (e.g., cdn.nytimes.com).
Certification Authority (CA). Publishers and CDNs can obtain signed public-key certificates from a certification authority
to bind a domain name (e.g., nytimes.com) to a particular
public key (0xfa315. . . ). When a CA signs a public-key
certificate, the CA is asserting that the same entity that controls
the domain named in the certificate also controls the public key
embedded in the certificate. Stickler only relies on HTTPS,
and therefore on CAs, for reliably obtaining the publisher’s
public key.
End Users. End users access the publisher’s website via their
web browser, which fetches assets from both the publisher’s
and the CDN’s servers.
B. Benefits and Risks of CDNs
Content distribution networks (CDNs) provide web content
caching as a third-party service for website publishers. CDN
providers typically have edge caching servers in multiple
locations around the world so that their caches are situated
in geographic proximity to a site’s end users. When an end
user makes a web request, the CDN serves the request from
its cache, otherwise it forwards the request to the publisher’s
server, caches the response, and returns the response to the
end user.
A website publisher may choose to employ a CDN for a
variety of reasons:
• CDNs can serve the bulk of a site’s assets from a longlived cache, dramatically reducing load on the publisher’s
servers.
• CDNs maintain edge servers around the world, so they
can service cached content to clients with relatively low
end-to-end latency.
• CDNs allow publishers to maintain availability in the face
of a rapid spike in traffic (i.e., a “flash crowd”) [14].
Assets cached by the CDN are served directly from the
CDN’s servers. As a result, if the end user fetches assets over
HTTPS, the HTTPS connection must terminate at the CDN.
The end user thus cannot distinguish between a well-behaved
CDN, which caches content from the website reliably and
serves it unmodified, and a malicious CDN that modifies the
publisher’s content.
Often, CDN-cached assets are served from a CDNcontrolled domain (e.g. nytimes.procdn.biz). However, publishers may choose to have the CDN cache content on a particular
subdomain (e.g., procdn.nytimes.com). Alternatively, a publisher may point the DNS records for her site to the CDN’s
servers to allow the CDN to serve as a front-end proxy for all
requests to the site. Either way, once the website has directed
the end user to retrieve content from a CDN, there is no way
for the the user to verify the authenticity of that content.
III.
T HREAT M ODEL AND S ECURITY G OALS
The goal of Stickler is to provide end-to-end integrity
protection for content served to clients through a CDN. In
particular, if the publisher correctly deploys Stickler, a client
will only execute JavaScript code signed by the publisher and
will only load publisher-signed content into the DOM, even in
the presence of active attacks by the CDN. We do not trust the
CDN for integrity. In particular, we assume that the the CDN
may:
• serve stale content from the cache instead of fetching a
newer version from the website.
• inject malicious client-side code into proxied responses
(to sniff passwords or track clients).
• modify page content (to include advertisements).
• downsample media files (to save bandwidth)
• respond in arbitrarily malicious ways to client requests.
Providing strong integrity guarantees for CDN-hosted content is critical because integrity and confidentiality are closely
linked. If the CDN can, for example, insert JavaScript code
into HTML pages served to web clients, the CDN can read
and exfiltrate passwords and other secret data via the client’s
DOM. By providing integrity guarantees, Stickler allows us to
protect against confidentiality attacks that a CDN can mount
by serving maliciously modified assets to the client.
We do, however, trust the CDN for availability: Stickler
does not attempt to maintain availability when a CDN refuses
to serve a publisher’s content to users at all. Since CDNs have
the site’s manifest file. Upon receiving the site manifest, the
bootloader script first checks that the manifest carries a valid
signature by the publisher, and then the bootloader executes
the manifest file as JavaScript.
(a) The client first fetches the site’s
index page.
(b) The site server returns the Stickler bootloader code.
(c) The bootloader script requests
the manifest via XHR.
(d) The CDN serves the signed
manifest file.
(e) Executing the manifest causes
the client to request the rest of the
assets.
(f) The CDN serves the signed site
assets (images, videos, etc.) to the
client.
Fig. 1.
The bootstrap process executed when a client visits a Sticklerprotected website.
a profit motive to provide availability (though not necessarily
integrity), we argue this model approximates the behavior of
“consumer-grade” CDNs.
Stickler does not attempt to hide the web client’s access
patterns from the CDN. Protecting access patterns is an important, though orthogonal, problem [15]–[18]. We also do not
attempt to protect against malicious CAs, though a publisher
could deploy Stickler in parallel with other defenses to protect
against CA compromise [19]–[21].
With Stickler we demonstrate that publishers can benefit
from the scalability and availability benefits of a CDN without
the integrity (and thus confidentiality) risks of serving assets
from a third-party server.
IV.
S YSTEM A RCHITECTURE
In this section, we step through the process of loading
a Stickler-protected website before examining each of the
architectural components in detail.
A. Design Overview
When a client browses to a Stickler-protected website, the
DNS record for the requested domain points the client to a
web server controlled by the publisher, not by the CDN. The
client makes an HTTPS connection to this server and requests
the index page for the website.
The publisher’s web server returns an HTML page with the
Stickler bootloader script embedded. This script contains (1)
the publisher’s public signature verification key, (2) JavaScript
code to download and verify the site’s assets, and (3) the
location of the site’s manifest file. Since this first request is
made over an HTTPS-authenticated connection directly to the
publisher’s server, authenticating the connection is sufficient
to authenticate the content. The publisher never needs to share
the private-key it uses to authenticate this initial connection.
When the client executes the bootloader script, the script
initiates an XHR request (i.e., AJAX request) to the CDN for
Executing the manifest file causes the client to generate a
series of XHR requests to the CDN for the rest of the site’s
assets, each of which bears a digital signature by the publisher.
When the CDN serves these assets to the client, the bootloader
script verifies the publisher’s signature on each asset and then
processes it by invoking a function defined in the manifest.
Typically, this processing just involves inserting the object into
a pre-specified location in the DOM. When the user interacts
with the site (e.g., by clicking a link), this could trigger more
remote asset loads and signature verifications as needed to
update the page content.
The publisher’s server only needs to serve the initial bootloader script—the client can request all of the site’s other assets
directly from the CDN—and the bootloader only changes if
the secret key changes, so it can be cached on the client.
This means that even though requests must go directly to
the publisher’s server, it will only impact load times the first
time a user visits the site, and should have little impact on
the publisher’s overall bandwidth usage. For example, in our
implementation, the gzipped bootloader is 1.2 KB, meaning
that a million unique visitors would require only 1.2 GB of
bandwidth. At the time of writing, transferring this amount of
outbound traffic from an Amazon EC2 instance costs less than
$0.11 [22].
B. Content Authentication
When the Stickler bootloader downloads an asset from the
CDN, the bootloader must verify the downloaded asset (to
prevent the CDN from maliciously modifying the asset). The
publisher has two options for each asset. An asset can either
be signed with the publisher’s private key or a cryptographic
hash of its contents can be embedded in the manifest. It is up
to the publisher to decide which to choose.
In general, a digital signature is useful if the content
is likely to change, since (1) the manifest (or whichever
JavaScript file loads the asset) need not change and (2) multiple
assets can exist concurrently if the CDN might serve the old
(but still valid) cached assets. On the other hand, embedding
the hash is suitable for content that is unlikely to ever change
and where changing the asset is difficult or infeasible. For
example, using a cryptographic hash to verify an asset allows
it to reside on a server the publisher has no control over,
such as a JavaScript or CSS library served from a third-party
(e.g. Bootstrap, jQuery). Our performance evaluation (Section VI) demonstrates that there is no significant client-side
performance difference between using cryptographic hashes or
digital signatures for verification.
To facilitate this, Stickler requires a modification to the
website authoring process. When the publisher builds their site,
they must digitally sign each asset served to the client. The
digital signing process can happen “on the fly” at page-load
time for dynamic portions of the site’s content, and can happen
at “compile time” for static portions of the site’s content. The
requirement to sign assets is not as burdensome as it may
sound: many publishers already run their HTML, JS, and CSS
files through minifiers and compression tools as part of their
“asset pipeline.” Adding a digital signing phase to this pipeline
would be relatively straightforward.
A malicious CDN might try to mount a replay attack by
serving stale but correctly signed content to the client. For
example, if a publisher discovers a security bug in one of its
site’s JavaScript files, the publisher would update the script
file to fix the bug. Instead of serving the new script file to
clients, a malicious CDN could continue to serve the validly
signed but out-of-date buggy file to clients. In this way, the
CDN could use a replay attack to cause a security bug in the
publisher’s site to persist.
To prevent replay attacks, the publisher affixes an “expiration date” to every asset before signing the asset and publishing
it to the CDN. After the Stickler bootloader downloads an asset
from the CDN and verifies the signature on the asset, it verifies
that the asset’s expiration date is in the future. By specifying
short expiration dates, the publisher gains more protection from
replay attacks, but must handle more traffic to their server as
assets expire. This is similar to choosing a cache expiration
time with the HTTP Expires header and, in most cases, we
expect publishers to choose the same value for both.
C. Fetching the Bootloader
As described above, the first time a client visits a Sticklerprotected web site, the client connects directly to the publisher’s server over HTTPS and downloads the Stickler bootloader. Since the bootloader contains the public signature
verification key used to validate subsequently loaded assets,
it is important that the publisher’s own servers serve the
bootloader to clients over an authenticated HTTPS connection.
As a result, the publisher must run a web server capable
of serving the bootloader at least once to every user. This
requirement is not necessarily prohibitive, though. The bootloader file is small (1.2 KB compressed) and can be cached
indefinitely, since it only needs to change if (a) there is a bug
in the bootloader or (b) the publisher wants to change their
public key.
There are several ways a publisher may force the client to
update a cached bootloader:
1) The publisher can serve the bootloader using the
Cache-Control: no-cache
HTTP
header.
Browsers still cache the content of the bootloader,
but check with the publisher’s server if it has changed
every time the user loads the web site. In this case,
the publisher’s server must be able to handle a request
every time each user visits the site, although unless the
bootloader has changed, the response will be small (an
empty 304 Not Modified response).
2) The publisher can instruct the browser (via a script in
the manifest file) to clear the cached bootloader using the window.location.reload(true) method.
This avoids making frivolous requests to the server when
the bootloader hasn’t changed, but won’t work if the
public key has changed.
3) The publisher can simply embed a new public key
in the bootloader and re-sign all assets with that key.
When a user first fetches an asset signed by the
new key, their cached bootloader will fail to verify
the asset. The bootloader may refresh itself (using
window.location.reload(true)) to check if the
publisher has embedded a new key before failing. In our
prototype, we chose to implement this strategy.
D. The Manifest File
The first file that the bootloader downloads is the site’s
manifest file, which contains a list of the site’s assets. For
each asset, the manifest file specifies:
• the asset’s URL,
• a hash of the asset’s contents (if using hash-based verification),
• an optional expiration date (to prevent the CDN from
serving stale assets),
• a callback function specifying what to do with the asset
once the bootloader has fetched and verified it (evaluate
it as JavaScript, place its contents in the DOM, etc.), and
• a callback function specifying what to do if the asset
integrity check fails.
The top-level manifest file that the bootloader fetches on
site load need not specify specify every asset that the client
might need. Instead, Stickler exposes a JavaScript API to the
client-side application allowing it to trigger subsequent remote
asset loads programmatically.
The Stickler API also allows the assets loaded in the toplevel manifest file to be manifest files themselves—allowing
for recursive manifest loads. This feature might be useful when
different web developers are responsible for different portions
of a web site—each developer could separately publish the
manifest for their own portion of a site without needing to
merge their changes into a top-level manifest file. Of course,
to minimize the number of round-trips to the CDN servers,
the application developers should minimize the depth of such
a “manifest tree,” since the number of round-trips increases
linearly with the depth of the manifest tree.
E. Dynamic Sites
Stickler is compatible with dynamically generated websites. There are two different methods a Stickler client-side
application could use to access dynamic content.
The first method is to fetch dynamic content directly
from the publisher’s server over HTTPS, bypassing the CDN
entirely. Since the publisher’s server will be hosted on a separate domain from the CDN’s servers anyways, the client can
connect directly to this domain to download dynamic content.
The bootloader need not verify the publisher’s signature on
assets fetched directly from the publisher’s server. In this case,
the remote peer and the content author are the same principal,
so HTTPS connection-level authentication is sufficient.
The second method is to fetch dynamically generated assets
via the CDN. This method might be useful if many clients
will request the same dynamic asset and the CDN can cache
the asset across client requests. In this case, the bootloader
does need to verify the publisher’s signature on the asset to
prevent the CDN from tampering with the asset in transit. To
allow for this integrity protection, the server-side application
producing the dynamic asset simply must bundle the asset as
a digitally signed blob that the bootloader can decode. The
publisher can implement this signing-and-packaging process as
part of its dynamic asset generation pipeline. The downsides
of this approach are that it requires the publisher’s server to
digitally sign every generated object and that it requires the
publisher to store its secret key online, where it may be less
secure.
F. Limitations
Stickler has a number of important drawbacks. First, Stickler requires the publisher to sign every asset served to the
client via a CDN. For some publishers, this may be relatively
straightforward, but for sites with a very large number of preexisting assets, signing every asset may be infeasible.
Second, Stickler imposes a performance penalty on the
user. When visiting a Stickler-protected site, the user’s browser
may have to perform a large number of signature verification
operations. Moreover, the user has to make an extra round-trip
to the CDN to get the manifest file. Although we argue that
these performance costs will not have a major impact on user
experience, they may be unacceptable for some publishers.
Section VI specifically evaluates the performance impact of
using Stickler.
Finally, Stickler prevents (by design) a CDN from minifying script files, shrinking images, or performing other sorts of
lossy compression on files without explicit publisher intervention. A publisher that heavily relies on their CDN to optimize
media files and static assets values the benefit of this CDNprovided service over the risks of trusting the CDN for content
integrity. However, it is important to note that Stickler does not
interfere with normal HTTP gzip compression, since the CDN
sees the content unencrypted.
V.
I MPLEMENTATION
To demonstrate that Stickler is a practical platform for
building full-featured websites, we have implemented Stickler
and developed a Stickler-protected demonstration website:
sticklerjs.org. There are two components to our Stickler implementation. The first component, the build framework, runs on
the system used to author the website. The second component,
the Stickler bootloader, runs on the client’s web browser.
Build Framework. The build framework runs on the development machines used to build the Stickler-protected website.
The build framework takes as input (1) the publisher’s secret
signing key and (2) the site’s static assets. The output of
the build process is a set of binary blobs that the publisher
serves to end users via the CDN. Each blob contains a single
site asset and is either signed with the publisher’s secret
signing key or its hash is embedded in the site’s manifest
file. Our implementation allows the publisher to specify, for
each asset, whether Stickler should use the signature- or hashbased technique to verify the asset’s provenance. The author
can additionally augment the manifest with assets from thirdparty sites (e.g. the jQuery CDN [23]) by specifying the URL
and the expected hash of the asset in the manifest.
We implemented the build framework as a set of
node.js [24] scripts. The framework outputs one data blob
per static site asset and each blob consists of signature on the
payload (the asset’s binary bytes, URL path and expiration) and
the payload itself. The build framework signs every blob, to
allow switching between hash- and signature-based verification
without rebuilding the asset blob files. Sites that already use a
web framework (e.g., Django or Ruby-on-Rails) for authoring
content could integrate our asset-signing process into the build
pipeline used in those tools.
Stickler uses OpenSSL [25] to generate signing keys. A
publisher may use any signing algorithm supported by both
OpenSSL and the WebCrypto implementation. In our prototype
and evaluations, we use the RSA signature algorithm with
2048-bit keys and RSA exponent e=65537 (the default for
OpenSSL). Using an RSA public key with e=3 would decrease
signature verification time, but since the cost of signature
verification is small compared with the overall costs of Stickler,
we chose to use OpenSSL’s default RSA exponent instead. For
both digital signature and hash verification we use the SHA256 hash function.
Stickler Bootloader. The Stickler bootloader is the HTML
document served to the web client when it makes the first
request to the publisher’s server. The bootloader HTML contains a copy of the publisher’s public signature verification key
and also contains JavaScript that loads the rest of the site. To
ensure the integrity of this file, Stickler requires that this asset
be served directly from the publisher’s server, rather than the
CDN. The CDN can serve all other site assets.
When executed, the bootloader’s JavaScript:
1) fetches the site manifest blob file from the CDN using a
cross-origin XHR request,
2) extracts the signature and payload from the blob,
3) uses the WebCrypto API [26] (or a JavaScript crypto
library like SJCL [27] if WebCrypto is unsupported by
the browser) to verify that the signature is valid, and
4) executes the contents of the manifest file using eval().
Since the bootloader is the core of Stickler, we explain each
of these steps in detail. In Step 1, the bootloader fetches the site
manifest file from a URL hard-coded into the bootloader. To
minimize the load on the publisher’s server, the CDN hosts the
manifest. Since Stickler-protected sites use a different domain
for CDN-hosted content than for publisher-hosted content, this
request will be a cross-origin XHR request. Although browsers
prohibit such cross-origin requests by default, the website
configures their assets to be served by the CDN using the
CORS headers, which instruct the client’s browser to allow the
request [28]. Most CDNs relay any HTTP headers directly to
the publisher’s servers, so the publisher must simply configure
its servers to serve cacheable assets with the appropriate CORS
header.
In Step 3, the bootloader script verifies the publisher’s
digital signature on the payload. The script verifies the signature against a public signature verification key hard-coded
into the bootloader script. If the verification check passes,
the bootloader executes the content of the manifest file. If it
fails, the bootloader first refreshes itself in case the publisher’s
verification key has changed, then shows an error to the enduser. The digital signature on the manifest file prevents a
malicious CDN (or man-in-the-middle) from tricking the web
client into executing arbitrary JavaScript.
Time (ms)
80
SHA-256
70
RSA/SHA-256
60
Base64
50
40
30
20
10
0
0
0.5
1
1.5
2
2.5
3
3.5
4
4.5
5
Asset size (MB)
Fig. 2. The top two lines represent the time to verify an asset using digital
signatures (RSA/SHA-256) and cryptographic hashes (SHA-256) in Firefox
depending on the size of the asset. The difference in performance between
the two methods was not statistically significant and was roughly 11 ms per
1 MB in asset size for each. The bottom line represents the additional time
to encode the asset as a Base64 data-uri if it is an image or video, in order
to insert it into the DOM. All experiments were run on a 3.5 Ghz quad-core
Intel i7 running Firefox 35.
In Step 4, the bootloader executes the manifest file. Since
the manifest file is just a JavaScript code snippet, the manifest
file can itself fetch and verify site assets, which can themselves
recursively fetch and verify other scripts and static resources.
Our entire bootloader is 104 lines of HTML and JavaScript
code and is 1.2 KB in size when compressed. The bootloader
code need not change often since the only two variables it
contains are (1) the publisher’s public key, and (2) the URL
of the manifest file.
VI.
P ERFORMANCE E VALUATION
The latency of loading an asset with Stickler is governed
by three factors: network latency, cryptographic overhead
(signature and hash operations), and the cost of loading the
asset into the DOM. Since network latency is not particular to
Stickler, our evaluation focuses on the last two factors.
Figure 2 compares the time required to perform an RSA2048 signature verification with the time required to perform
a SHA-256 hash verification. Our experiments show that for
small assets, both signature and hash verification complete in
under 1 ms. For larger assets, traversing the contents dominates
performance and there is no measurable difference between
signing and hashing. Each additional 1 MB in asset size
introduces an extra 11 ms overhead for both signature and hash
verification. In general, verification in either method constitutes
about a 1% overhead compared to downloading that asset over
a 10 Mbps network connection.
Images and other media need to further be converted to
Base64 data-uris as an artifact of how they are embedded
in the DOM. Figure 2 shows a similar experiment (using the
same assets) where binary data was converted to data-uris.
The results show that this conversion, too, is relatively small.
Each additional 1 MB in asset size requires an extra 5 ms to
convert it to Base64. As a concrete example, downloading a
5 MB image over a 10 Mbps connection takes approximately
4 seconds, while verifying its signature or hash and converting
it to Base64 takes less than 80 ms in total.
Figure 3 compares the time it takes to render a Sticklerprotected page with the time it takes to load an unprotected
static HTML page as the number of assets on the page
varies. In the experiment, we loaded a page with a varying
number of 100 KB images (between 1 and 96) over the local
network. The experiments ran on an Intel W3565 quad-core
based workstation with hyperthreading running Firefox 35
and Chrome 40, and on the Google Nexus 5 mobile phone
running mobile Firefox 35. As the figure demonstrates, the
page render time increases roughly linearly with the number
of images. Even with a relatively content-heavy page (10 MB
of media assets), the page renders within 1 second on the
Chrome browser and within 6 seconds on a mobile phone
running Firefox. While the performance with and without
Stickler on Chrome is comparable, Stickler imposes roughly a
5× performance penalty when using Firefox when the number
of assets on the page is large. Whether or not this overhead is
acceptable will depend on the particular security requirements
of the application, but we expect that for especially sensitive
sites (e.g., a health data site), publishers will be willing to pay
a performance cost for a security benefit.
VII.
R ELATED W ORK
The idea of using a small JavaScript bootloader to verify
the provenance of subsequently downloaded assets has been
in the Web folklore for some time and is even deployed on
the Web [29], [30]. The goal of this paper is to formalize the
threat model and security guarantees that such a system can
provide and evaluate the performance costs of using such an
architecture without modifying existing browsers.
An important piece of related work is the W3C Subresource
integrity (SRI) working draft, which would augment browsers
with tools to provide integrity guarantees in the face of a
malicious CDN [9]. When serving a page to an SRI-enabled
browser, the site author can specify the expected “fingerprint”
(i.e., SHA-256 hash) of the resource to be loaded. When the
browser fetches the resource, it checks the value against the
expected fingerprint and only renders the object if the actual
and expected fingerprints match. Browsers can enforce similar
integrity checks using the “script hash” feature of Content
Security Policy 2 (CSP) W3C candidate recommendation [13].
Stickler differs from these mechanisms in two important
ways. First, Stickler protects content integrity without modifying the browser. Not needing to modify the browser makes
Stickler useful as an interim solution until modern browsers
support these features and as a long-term solution for browsers
that may never support these features (e.g., on embedded
devices). Second, Stickler allows the publisher to use digital
signatures to verify assets, in addition to cryptographic hashes.
When using digital signatures to protect content integrity, the
publisher need only update the Stickler bootloader when it
changes its public key. In contrast, if a publisher used only SRI
to protect page integrity, the publisher would need to update the
cryptographic hash in the bootloader (and potentially in many
other script files as well) every time it changed any piece of
content on a page. Use of digital signatures streamlines the
page update process and moves full-site integrity protection
towards practicality.
That said, it would be possible to augment the SRI and CSP
recommendations to support signature-based asset integrity
checks in the browser as well. One possible way to implement
Firefox
Chrome
Firefox Mobile
1
3
2
1
0
12
Stickler
No Stickler
0.8
Render time (s)
Stickler
No Stickler
4
Render time (s)
Render time (s)
5
0.6
0.4
0.2
0
0
1
2
3
4
5
6
7
Total asset size (MB)
8
9
10
Stickler
No Stickler
10
8
6
4
2
0
0
1
2
3
4
5
6
7
8
9
10
0
Total asset size (MB)
1
2
3
4
5
6
7
8
9
10
Total asset size (MB)
Fig. 3. Stickler imposes roughly a 5× performance overhead on Firefox, a 2× overhead on Chrome, and negligible overhead on a mobile device. In these
experiments we loaded web pages into each browser while varying the number of images on the page. We measured the total time, including network latency
to a server on the local host, to load the entire page. We ran the experiments on a quad-core Intel W3656 and a Google Nexus 5 mobile phone.
this feature would be to have the initial HTTPS request to the
publisher’s servers return an HTTP header indicating the public
key that the publisher will use to sign assets served over the
CDN. When the browser makes subsequent requests for site
assets (images, JS files, etc.) from the CDN’s servers, it would
require that each response be signed with the publisher’s asset
signing key. The CDN could return these signatures to the end
user’s browser in a special HTTP response header. Although
this appears to be a promising direction for content integrity
protection, we leave a full security analysis of this scheme to
future work.
Liang et al. describe the weaknesses of today’s “man-inthe-middle” model for using HTTPS with CDNs [31]. They
propose allowing publishers to specify (in their DNS records) a
set of SSL certificates that a client should accept when fetching
content from the publisher’s CDN. This mechanism makes it
possible for the publisher to later revoke a CDN’s certificate
without having to revoke its own certificate. Unlike Stickler,
the solution of Liang et al. does not protect the integrity of
content served by a publisher’s CDN.
Prior research on CDNs has focused on the design of such
systems and their effectiveness at maintaining availability in
the presence of “flash crowds” (traffic spikes) [3]–[8]. The
Firecoral [32] system uses a browser extension to verify the
integrity of content served from a peer-to-peer CDN whereas
Stickler does not require modifying the browser.
Other prior work aims to protect against common browserrelated security pitfalls. Bonneau’s S-links make it possible
to provide integrity guarantees in the presence of malicious
certification authorities (CAs) [19]. Erlingsson et al. propose
a system for protecting against code-injection attacks by modifying the browser to read a special server-provided security
policy before processing JavaScript in the page [33]. Both of
these techniques require modifying the browser and neither
provides end-to-end integrity when used with a CDN.
VIII.
Stickler is a framework for protecting the integrity of a
web application from a faulty or malicious CDN. Stickler
uses a JavaScript-based bootloader delivered directly from the
publisher’s domain to verify the provenance of site assets
delivered via a minimally trusted CDN. Crucially, Stickler
does not require browser extensions or modifications and it is
compatible with popular web publishing tools and techniques.
Our implementation and evaluation of Stickler demonstrate its
practicality and performance.
With this work, we show that website publishers can
reap the manifold performance and cost benefits of using a
CDN without having to put unnecessary trust in the CDN’s
correctness or honesty.
ACKNOWLEDGMENTS
We would like to thank the anonymous reviewers for their
thoughtful comments. David Terei, Deian Stefan, and Keith
Winstein all provided helpful suggestions on the content and
presentation of the paper and, on occasion, politely listened to
us rant about CDNs. This work was supported by an NDSEG
fellowship, NSF, DARPA, an IARPA project provided via
DoI/NBC, a grant from ONR, and by a Google faculty scholarship. Opinions, findings and conclusions or recommendations
expressed in this material are those of the author(s) and do not
necessarily reflect the views of DARPA or IARPA.
R EFERENCES
[1]
[2]
[3]
[4]
[5]
Mylar [34] protects the confidentiality of data stored on
an untrusted application server by encrypting it in the client’s
browser. Mylar uses a browser extension to validate the clientside code before executing it. Since we aim only to protect
against a malicious CDN (not a malicious application server),
we can avoid the use of a browser extension. Christodorescu
discussed the value of using client-side cryptography to protect
against server compromise [35].
C ONCLUSION
[6]
[7]
[8]
CloudFlare, “Who’s using CloudFlare?” https://www.cloudflare.com/
customers. Accessed on 12 March 2015.
M. Prince, “CloudFlare blog: Introducing Universal SSL,” https://blog.
cloudflare.com/introducing-universal-ssl/, September 2014.
M. J. Freedman, E. Freudenthal, and D. Mazieres, “Democratizing
content publication with coral.” in NSDI, vol. 4, 2004, pp. 18–18.
M. J. Freedman, “Experiences with CoralCDN: A five-year operational
view,” in NSDI, 2010, pp. 95–110.
C. Huang, A. Wang, J. Li, and K. W. Ross, “Measuring and evaluating
large-scale CDNs,” in IMC. ACM, October 2008.
B. Livshits and D. Molnar, “Empowering browser security for mobile
devices using smart CDNs,” in W2SP, 2010.
G. Pallis and A. Vakali, “Insight and perspectives for content delivery
networks,” Communications of the ACM, vol. 49, no. 1, pp. 101–106,
2006.
L. Wang, K. Park, R. Pang, V. S. Pai, and L. L. Peterson, “Reliability
and security in the CoDeeN content distribution network.” in USENIX
ATC, 2004, pp. 171–184.
[9]
[10]
[11]
[12]
[13]
[14]
[15]
[16]
[17]
[18]
[19]
[20]
“Subresource integrity: W3C first public working draft,” http://www.
w3.org/TR/SRI/, March 2014.
S. Blake-Wilson, M. Nystrom, D. Hopwood, J. Mikkelsen, and
T. Wright, “Transport layer security (tls) extensions,” http://www.
rfc-editor.org/info/rfc3546, June 2003.
Android Issue Tracker, “Issue 12908: Https websites that support server
name indication (sni) don’t work,” https://code.google.com/p/android/
issues/detail?id=12908#c13, July 2011.
E.
Law,
“Understanding
certificate
name
mismatches,”
http://blogs.msdn.com/b/ieinternals/archive/2009/12/07/
certificate-name-mismatch-warnings-and-server-name-indication.aspx,
December 2009.
“Content security policy level 2: W3C candidate recommendation,” http:
//www.w3.org/TR/CSP2/, Febuary 2015.
P. Wendell and M. J. Freedman, “Going viral: flash crowds in an open
cdn,” in IMC. ACM, 2011, pp. 549–558.
D. Chaum, “Untraceable electronic mail, return addresses, and digital
pseudonyms,” CACM, vol. 24, no. 2, pp. 84–90, 1981.
——, “The Dining Cryptographers Problem: Unconditional sender and
recipient untraceability,” Journal of Cryptology, vol. 1, no. 1, pp. 65–75,
1988.
B. Chor, E. Kushilevitz, O. Goldreich, and M. Sudan, “Private information retrieval,” JACM, vol. 45, no. 6, pp. 965–981, 1998.
R. Dingledine, N. Mathewson, and P. Syverson, “Tor: The secondgeneration onion router,” in USENIX Security, August 2004.
J. Bonneau, “S-links: Why distributed security policy requires secure
introduction,” in W2SP, May 2013.
C. Evans and C. Palmer, “Public key pinning extension for HTTP,”
IETF Web Security Internet-Draft, November 2011.
[21]
B. Laurie, A. Langley, and E. Kasper, “Certificate transparency,” IETF
Network Working Group Internet-Draft, June 2013.
[22]
“Amazon EC2 Pricing,” https://aws.amazon.com/ec2/pricing/.
[23]
“jQuery,” http://www.jquery.com/.
[24]
“node.js,” http://www.nodejs.org/.
[25]
“Openssl,” https://www.openssl.org/.
[26]
“Web cryptography API: W3C candidate recommendation,” http://www.
w3.org/TR/WebCryptoAPI/, December 2014.
[27]
E. Stark, M. Hamburg, and D. Boneh, “Stanford javascript crypto
library,” http://bitwiseshiftleft.github.io/sjcl/.
[28]
“Cross-origin resource sharing: W3C recommendation,” http://www.w3.
org/TR/cors/, January 2014.
[29]
“MEGA,” https://mega.co.nz/.
[30]
“PeerCDN,” https://peercdn.com/.
[31]
J. Liang, J. Jiang, H. Duan, K. Li, T. Wan, and J. Wu, “When HTTPS
meets CDN: A case of authentication in delegated service,” in Security
and Privacy. IEEE, 2014, pp. 67–82.
[32]
J. Terrace, H. Laidlaw, H. E. Liu, S. Stern, and M. J. Freedman,
“Bringing P2P to the web: security and privacy in the Firecoral
network.” in IPTPS, 2009, p. 7.
[33]
U. Erlingsson, V. B. Livshits, and Y. Xie, “End-to-end web application
security.” in HotOS, 2007.
[34]
R. A. Popa, E. Stark, J. Helfer, S. Valdez, N. Zeldovich, M. F. Kaashoek,
and H. Balakrishnan, “Building web applications on top of encrypted
data using Mylar,” in NSDI, April 2014.
[35] M. Christodorescu, “Private use of untrusted web servers via opportunistic encryption,” in W2SP, May 2008.